Why Reporting the Components Matters
A pooled estimate alone does not show how uncertain the result is, how much that uncertainty reflects differences among imputations, or what reference distribution was used for inference. A transparent report therefore gives the estimate together with its pooled standard error, degrees of freedom, confidence interval, and p-value when a hypothesis test is relevant. Reporting selected variance components can make the source of uncertainty clearer.
The formulas below concern a single scalar estimand on the same scale in every completed data set. The completed-data estimates and their variance estimates must be valid for that estimand; imputations must reflect the uncertainty required for multiple-imputation inference; and the completed-data estimates should be sufficiently well approximated by a normal sampling distribution for a t-based summary to be appropriate. These conditions do not repair a poorly specified imputation or analysis model. In particular, the estimates must not change meaning or scale across imputations.
The within-imputation variance \(\bar U\) summarizes sampling uncertainty within completed data sets. The between-imputation variance \(B\) summarizes how much the estimates differ across them. The multiplier \(1+1/m\) in \(T\) includes both the between-imputation contribution and the finite number of imputations. It is part of the total-variance formula, not an optional adjustment.
From Variance Components to Inference
For the usual large-complete-data-sample Rubin approximation, define the relative increase in variance due to nonresponse as \(r=(1+1/m)B/\bar U\). The reference degrees of freedom are \(\nu_{\text{old}}=(m-1)(1+1/r)^2\). Under the null hypothesis \(H_0:Q=Q_0\), the pooled test statistic is \(t=(\bar Q-Q_0)/SE\), compared with a t distribution having \(\nu_{\text{old}}\) degrees of freedom. For a two-sided test, the p-value is \(2P(t_{\nu_{\text{old}}}\geq |t|)\). A \(100(1-\alpha)\%\) confidence interval is \(\bar Q\pm t_{1-\alpha/2,\nu_{\text{old}}}SE\).
If the complete-data analysis has finite degrees of freedom \(\nu_{\text{com}}\), a small-sample adjustment may be needed. One commonly used Barnard–Rubin adjustment defines \(\gamma=(r+2/(\nu_{\text{old}}+3))/(r+1)\), then calculates \(\nu_{\text{obs}}=((\nu_{\text{com}}+1)/(\nu_{\text{com}}+3))\nu_{\text{com}}(1-\gamma)\), and combines the two degrees of freedom as \(\nu=(\nu_{\text{old}}\nu_{\text{obs}})/(\nu_{\text{old}}+\nu_{\text{obs}})\). Use the adjusted \(\nu\) for the t-based interval and test when that adjustment is appropriate, and identify it in the report. The degrees of freedom are part of the inferential method, not merely a display detail.
Worked Example: A Pooled Estimate and Its Confidence Interval
Suppose five hypothetical imputations give estimates of a treatment contrast, measured in points, of \(1.8,\ 2.0,\ 2.2,\ 2.4,\ 2.1\). The corresponding estimated variances, in points squared, are \(0.04,\ 0.05,\ 0.04,\ 0.06,\ 0.05\). Assume each completed-data variance estimate is appropriate and the estimates refer to the same contrast and scale.
The pooled estimate is \(\bar Q=(1.8+2.0+2.2+2.4+2.1)/5=10.5/5=2.10\) points. The average within-imputation variance is \(\bar U=(0.04+0.05+0.04+0.06+0.05)/5=0.0480\) points squared.
The deviations from \(2.10\) are \(-0.30,-0.10,0.10,0.30,0\). Thus \(B=[0.09+0.01+0.01+0.09+0]/4=0.0500\) points squared. The total variance and pooled standard error are
Here \(r=(1.2)(0.0500)/0.0480=1.25\), so \(\nu_{\text{old}}=4(1+1/1.25)^2=4(1.8)^2=12.96\). For a 95% interval, \(t_{0.975,12.96}\) is approximately \(2.160\). The margin of error is \(2.160(0.328633\ldots)=0.7098\) points, giving \(2.10\pm0.7098\), or approximately \(1.390\) to \(2.810\) points. The interpretation is that this procedure gives intervals that cover the fixed contrast in 95% of repeated applications under its assumptions; it is not a 95% probability statement about this realized interval.
To test \(H_0:Q=0\), \(t=2.10/\sqrt{0.108}\approx6.390\). With 12.96 degrees of freedom, the two-sided p-value is less than \(0.001\). The data therefore provide strong evidence against a zero contrast under the stated model and pooling approximation.
Reporting Variance and Missing-Information Measures
A useful way to describe the variance decomposition is to report \(\bar U\), \(B\), and \(T\), alongside \(m\). Their units are squared units of the estimand; the pooled standard error has the estimand’s original units. Two dimensionless summaries are often helpful. The relative increase in variance \(r=(1+1/m)B/\bar U\) compares the between-imputation contribution, including its finite-\(m\) factor, with the within-imputation variance. Rubin’s large-sample fraction of missing information is \(\lambda=(1+1/m)B/T\), the fraction of total variance contributed by that same between-imputation term. These are related but not interchangeable quantities.
Some software reports a finite-degrees-of-freedom estimate of the fraction of missing information instead of, or in addition to, \(\lambda\). One common version is \(\gamma=(r+2/(\nu_{\text{old}}+3))/(r+1)\). State which measure is reported and how it is defined; a number labelled only “missing information” may be ambiguous. These measures describe uncertainty attributable to missingness under the analysis and imputation procedure. They are not direct estimates of the percentage of observations missing.
Worked Example: Interpreting the Variance Components
Use the preceding hypothetical results: \(m=5\), \(\bar U=0.0480\), \(B=0.0500\), and \(T=0.1080\), all variances in points squared. The between-imputation contribution included in \(T\) is \((1+1/m)B=1.2(0.0500)=0.0600\) points squared.
The relative increase is \(r=0.0600/0.0480=1.25\). The Rubin fraction is \(\lambda=0.0600/0.1080=0.5556\). Thus, in this variance decomposition, the between-imputation contribution is 55.56% of total variance; the remaining 44.44% is \(\bar U/T=0.0480/0.1080\). It would be incorrect to say that 55.56% of cases were missing, or that the estimates vary by 55.56%. The quantity is a ratio of variance components, not a missing-case rate or a measure in points.
Worked Example: A Finite-Complete-Degrees-of-Freedom Adjustment
Consider a separate hypothetical analysis with \(m=5\), \(\bar Q=0.600\), \(\bar U=0.0400\), and \(B=0.0100\). Suppose the complete-data analysis has \(\nu_{\text{com}}=20\) degrees of freedom. The estimates and variance estimates are assumed to refer to the same scalar contrast, and the completed-data variance estimates are assumed valid.
The total variance is \(T=0.0400+1.2(0.0100)=0.0520\), so \(SE=\sqrt{0.0520}=0.2280\). The relative increase is \(r=0.0120/0.0400=0.3000\), giving \(\nu_{\text{old}}=4(1+1/0.3000)^2=75.11\). The Rubin fraction is \(\lambda=0.0120/0.0520=0.2308\).
For the Barnard–Rubin adjustment, \(\gamma=(0.3000+2/(75.11+3))/(1+0.3000)\approx0.2505\), so \(\nu_{\text{obs}}=(21/23)(20)(1-0.2505)\approx13.69\), and \(\nu=(75.11)(13.69)/(75.11+13.69)\approx11.58\). Using \(t_{0.975,11.58}\approx2.188\), the 95% interval is \(0.600\pm2.188\sqrt{0.0520}\approx0.600\pm0.4989\), or approximately \(0.101\) to \(1.099\) units of the contrast. The test statistic for \(H_0:Q=0\) is \(0.600/\sqrt{0.0520}\approx2.631\); with 11.58 degrees of freedom, its two-sided p-value is approximately \(0.022\). The interval excludes zero, consistent with that two-sided test at the 5% level.
How to Present the Results Clearly
A compact report should make the estimate, uncertainty, and inferential choices recoverable. Include the estimand and its units, \(m\), \(\bar Q\), \(SE\), the degrees of freedom and adjustment used, the confidence level and interval, and the null value and sidedness for any p-value. If variance components are relevant, give \(\bar U\), \(B\), and \(T\), and identify whether a reported missing-information measure is \(\lambda\) or a finite-df version such as \(\gamma\). Round enough to preserve useful information, but do not imply more precision than the analysis supports; calculate intervals and test statistics from unrounded values before rounding the displayed results.
For example, the first worked result could be reported as: “Across five imputations, the pooled treatment contrast was 2.10 points (SE 0.329; 95% CI 1.390 to 2.810; \(\nu_{\text{old}}=12.96\); two-sided \(p<0.001\)). The within-imputation variance was 0.0480 points squared, the between-imputation variance was 0.0500 points squared, and the total variance was 0.1080 points squared.” This gives readers the result and enough information to understand how its uncertainty was summarized.
Confirm that all \(Q_i\) estimate the same quantity on the same scale, and that every \(U_i\) is a variance, not a standard error.
Compute \(\bar Q\), \(\bar U\), \(B\), \(T\), and \(SE=\sqrt T\), retaining unrounded values during the calculations.
Use the appropriate large-sample or finite-complete-data adjustment, then use that reference distribution consistently for intervals and tests.
Give the estimate and its units, uncertainty, inference, and any variance or missing-information measure with its definition.
Common Mistakes and Limitations
- Pooling standard errors instead of variances. The pooling formula combines \(U_i\), which are variances. Averaging \(\sqrt{U_i}\) and treating that as \(\sqrt T\) omits the between-imputation component and is not Rubin pooling.
- Omitting the finite-\(m\) factor. Replacing \(T=\bar U+(1+1/m)B\) by \(\bar U+B\) changes the pooled standard error and can change the interval and test.
- Reporting a p-value without its reference distribution. A pooled p-value requires the pooled statistic and its degrees of freedom. Averaging imputation-specific p-values is not a substitute.
- Calling \(\lambda\) the missing-data percentage. It is a variance fraction, not the proportion of records with missing values.
- Ignoring unstable components. With few imputations, \(B\), the degrees of freedom, and the resulting inference can be imprecisely estimated. An unexpectedly large or very small \(B\) should prompt checks of the estimates, imputation variability, and model behavior rather than automatic acceptance of the summary.
The pooled interval and test also rely on the imputation and analysis procedures being appropriate for the estimand. A precise-looking standard error cannot compensate for incompatible estimands, invalid completed-data variance estimates, or an imputation procedure that fails to represent relevant uncertainty. Reporting the components improves transparency, but it is not itself a diagnostic proof that those assumptions hold.
Check Your Understanding
Try each question, then compare your work with the solutions below.
- For \(m=4\), \(\bar U=0.10\), and \(B=0.02\), calculate \(T\), \(SE\), and \(r\).
- In one sentence, distinguish \(r\) from \(\lambda\).
- Why must a reported pooled confidence interval identify the degrees of freedom or adjustment used?
- A report says “the fraction of missing information was 40%, so 40% of participants had missing data.” What is wrong with this interpretation?
Solutions. (1) \(T=0.10+(1+1/4)(0.02)=0.125\), so \(SE=\sqrt{0.125}=0.3536\), and \(r=(1.25)(0.02)/0.10=0.2500\). (2) \(r\) compares the between-imputation variance contribution with \(\bar U\), whereas \(\lambda\) is that contribution as a fraction of total variance \(T\). (3) The degrees of freedom determine the t critical value and tail probability, so omitting them leaves the inferential calculation unclear and potentially unreproducible. (4) The fraction of missing information describes a variance decomposition under the analysis; it is not the percentage of participants with missing values.